Scholarlink

Multi-object tracking (MOT) in the incident video is an integral part of modern surveillance systems and security systems. It plays an important role in situational awareness, providing timely and effective responses by providing real-time data and insights. These video streams require continuous detection and monitoring, which is particularly challenging in complex dynamic environments such as traffic accidents, public meetings, and emergencies. The primary goal of MOT is to track every object precisely and consistently despite its numerous obstructions, connections, and variants. Advanced algorithms, including deep learning, data association techniques, and motion prediction, must be used to enhance tracking robustness and accuracy in event videos. This paper proposes a fully automated approach for multiple object detection and tracking in incident videos based on YOLO v9 and CNN-LSTM. The first stage of the proposed system in this paper is based on accurate multiple object detection using YOLO. YOLOv9 (You Only Look Once, Version 9) is used for automated multiple object detection. YOLO represents a major advance in real-time object recognition technology, building on its predecessor to deliver unprecedented speed and accuracy. The second stage is based on multiple object understanding using CNN-LSTM to generate accurate object image captions. CNN-LSTM (Convolutional Neural Network—Long Short-Term Memory) model of image reference generation is a powerful framework that combines the power of deep learning algorithms to generate annotations and information with and around the image’s description. Provides position. These features are then transferred to the LSTM network, which excels at sequential data processing, producing a sequence of words that are natural language subjects. This method carefully captures spatial information in the image and temporal information that depends on the draft text. The combination of CNNs and LSTMs makes the model more contextual and syntactically correct, making it more efficient for applications such as passive image registration, assistive technologies for visual impairment, and content-based image reception. Explains. The proposed approach performs well using a standard date set (Flicker). The experimental result shows high performance regarding the average accuracy, which exceeds 83.3%.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep MOT: Fully Automated Multiple Object Tracking in Incident Videos Using Deep Convolutional Neural Network-Based Short-Term Memory LSTM and YOLO v.9

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sara A. Alhashimi,
  • Adil Al-Azzawi

摘要

Multi-object tracking (MOT) in the incident video is an integral part of modern surveillance systems and security systems. It plays an important role in situational awareness, providing timely and effective responses by providing real-time data and insights. These video streams require continuous detection and monitoring, which is particularly challenging in complex dynamic environments such as traffic accidents, public meetings, and emergencies. The primary goal of MOT is to track every object precisely and consistently despite its numerous obstructions, connections, and variants. Advanced algorithms, including deep learning, data association techniques, and motion prediction, must be used to enhance tracking robustness and accuracy in event videos. This paper proposes a fully automated approach for multiple object detection and tracking in incident videos based on YOLO v9 and CNN-LSTM. The first stage of the proposed system in this paper is based on accurate multiple object detection using YOLO. YOLOv9 (You Only Look Once, Version 9) is used for automated multiple object detection. YOLO represents a major advance in real-time object recognition technology, building on its predecessor to deliver unprecedented speed and accuracy. The second stage is based on multiple object understanding using CNN-LSTM to generate accurate object image captions. CNN-LSTM (Convolutional Neural Network—Long Short-Term Memory) model of image reference generation is a powerful framework that combines the power of deep learning algorithms to generate annotations and information with and around the image’s description. Provides position. These features are then transferred to the LSTM network, which excels at sequential data processing, producing a sequence of words that are natural language subjects. This method carefully captures spatial information in the image and temporal information that depends on the draft text. The combination of CNNs and LSTMs makes the model more contextual and syntactically correct, making it more efficient for applications such as passive image registration, assistive technologies for visual impairment, and content-based image reception. Explains. The proposed approach performs well using a standard date set (Flicker). The experimental result shows high performance regarding the average accuracy, which exceeds 83.3%.